CARBayes ST

Spatio-Temporal Areal Unit Modelling in R with
Conditional Autoregressive Priors Using the

CARBayesST Package
Duncan Lee Alastair Rushworth Gary Napier

University of Glasgow University of Strathclyde University of Glasgow
Abstract
Spatial data relating to non-overlapping areal units are prevalent in fields such as
economics, environmental science, epidemiology and social science, and a large suite of
modelling tools have been developed for analysing these data. Many utilise conditional
autoregressive (CAR) priors to capture the spatial autocorrelation inherent in these data,
and software such as CARBayes and R-INLA have been developed to make these models
easily accessible to others. Such spatial data are typically available for multiple time
periods, and the development of methodology for capturing temporally changing spatial
dynamics is the focus of much current research. A sizeable proportion of this literature has
focused on extending CAR priors to the spatio-temporal domain, and this article presents
the R package CARBayesST, which is the first dedicated software for spatio-temporal
areal unit modelling with conditional autoregressive priors. The software can fit a range
of models focused on different aspects of space-time modelling, including estimation of
overall space and time trends, and the identification of clusters of areal units that exhibit
elevated values. This paper outlines the class of models that the software can implement,
before applying them to one simulated and two real examples in the fields of epidemiology
and housing market analysis.
Keywords: Bayesian inference, conditional autoregressive priors, R package, spatio-temporal

areal unit modelling.
1. Introduction
Areal unit data are a type of spatial data where the observations relate to a set of K con-
tiguous but non-overlapping areal units, such as electoral wards or census tracts. Each ob-
servation relates to an entire areal unit, and thus is typically a summary measure such as
an average, proportion, or total, for the entire unit. Examples include the total yield in
sectors in an agricultural field trial (Besag and Higdon 1999), the proportion of people who
are Catholic in lower super output areas in Northern Ireland (Lee, Minton, and Pryce 2015),
the average score on SAT college entrance exams across US states (Wall 2004), and the
total number of cases of chronic obstructive pulmonary disease from populations living in
counties in Georgia, USA (Choi and Lawson 2011). Areal unit data such as these have be-
come increasingly available in recent times, due to the creation of databases such as Scottish
Statistics (http://statistics.gov.scot/), Health and Social Care Information Centre In-
dicator Portal (http://www.hscic.gov.uk/indicatorportal), and cancer registries such as
2 CARBayesST: Spatio-Temporal Areal Unit Modelling
the Surveillance Epidemiology and End Results programme (http://seer.cancer.gov).
These databases provide data on a set of K areal units for N consecutive time periods,
yielding a rectangular array of K N spatio-temporal observations. The motivations for
modelling these data are varied, and include estimating the effect of a risk factor on a re-
sponse (see Wakefield 2007 and Lee, Ferguson, and Mitchell 2009), identifying clusters of
contiguous areal units that exhibit an elevated risk of disease compared with neighbouring
areas (see Charras-Garrido, Abrial, and de Goer 2012 and Anderson, Lee, and Dean 2014),
and quantifying the level of segregation in a city between two or more different groups (see
Lee et al. 2015). As a result different space-time structures have been proposed for modelling
spatio-temporal data, which depend on the underlying question(s) of interest and the goals
of the analysis.
However, a common challenge when modelling these data is that of spatio-temporal autocor-
relation, namely that observations from geographically close areal units and temporally close
time periods tend to have more similar values than units and time periods that are further
apart. Temporal autocorrelation occurs because the data relate to largely the same popu-
lations in consecutive time periods, while spatial autocorrelation can arise for a number of
reasons. The first is unmeasured confounding, which occurs when a spatially patterned risk
factor for the response variable is not included in a regression model, and hence its omission
induces spatial structure into the residuals. Other causes of spatial autocorrelation include
neighbourhood effects, where the behaviours of individuals in an areal unit are influenced
by individuals in adjacent units, and grouping effects where groups of people with similar
behaviours choose to live close together.
Predominantly, a Bayesian approach is taken to modelling these data in the statistical com-
munity, where the spatio-temporal structure is modelled via sets of autocorrelated random
effects. Conditional autoregressive (CAR, Besag, York, and Mollie 1991) priors and spatio-
temporal extensions thereof are typically assigned to these random effects to capture the
autocorrelation, which are special cases of a Gaussian Markov Random Field (GMRF). A
range of models have been proposed with different space-time structures, which can be used
to answer different questions of interest about the data. For example, Knorr-Held (2000)
proposed an ANOVA style decomposition of the data into overall spatial and temporal trends
and a space-time interaction, which allows common patterns such as the region wide average
temporal trend to be estimated. In contrast, Li, Best, Hansell, Ahmed, and Richardson (2012)
developed a model for identifying areas that exhibited unusual trends that were different from
the overall region wide trend, while Lee and Lawson (2016) presented a model for identifying
spatio-temporal clustering in the data.
An array of freely available software can now implement purely spatial areal unit models, rang-
ing from general purpose statistical modelling software such as BUGS (Lunn, Spiegelhalter,
Thomas, and Best 2009) and R-INLA (Rue, Martino, and Chopin 2009), to specialist spatial
modelling packages in the statistical software R (R Core Team 2016) such as CARBayes (Lee
2013), spatcounts (Schabenberger 2009) and spdep (Bivand and Piras 2015). However, due to
the flexibility of general purpose software, implementing spatial models, in say BUGS, requires
a degree of programming that is non-trivial for applied researchers. Specialist software for
Duncan Lee, Alastair Rushworth, Gary Napier 3
spatio-temporal modelling is much less well developed, with examples for geostatistical data
including spTimer (Bakar and Sahu 2015) and spBayes (Finley, Banerjee, and Gelfand 2015).
For areal unit data the surveillance (Paul and Meyer 2016) package models epidemic data,
the plm (Croissant and Millo 2008) and splm (Millo and Piras 2012) packages model panel
data, while the nlme (Pinheiro, Bates, DebRoy, Sarkar, and R Core Team 2015) and lme4
(Bates, Machler, Bolker, and Walker 2015) packages have functionality to model spatial and
temporal random effects structures. However, software to fit a range of spatio-temporal areal
unit models with CAR type autocorrelation structures is not available, which has motivated
us to develop the R package CARBayesST.
CARBayesST can fit models with different spatio-temporal structures, which allows the user
to answer different questions about their data within a common software environment. A
number of models can be implemented, including: (i) a spatially varying linear time trends
model (similar to Bernardinelli, Clayton, Pascutto, Montomoli, Ghislandi, and Songini 1995);
(ii) a spatial and temporal main effects and interaction model (similar to Knorr-Held 2000);
(iii) a spatially autocorrelated autoregressive time series model (Rushworth, Lee, and Mitchell
2014); (iv) a model with a common temporal trend but varying spatial surfaces (similar
to Napier, Lee, Robertson, Lawson, and Pollock 2016); (v) a spatially adaptive smoothing
model for localised spatial smoothing (Rushworth, Lee, and Sarran 2016); and (vi) a spatio-
temporal clustering model (Lee and Lawson 2016). The software has the same syntax and
feel as the R package CARBayes for spatial areal unit modelling, and retains all of its easy-to-
use features. These include specifying the spatial adjacency information via a single matrix
(unlike BUGS that requires 3 separate list objects), fitting models via a one-line function
call, and compatibility with CARBayes that allows it to share the latters model summary
functionality for interpreting the results. The models available in this software can be fitted
to binomial, Gaussian (in some cases) or Poisson data, and Section 2 summarises the models
that can be implemented. Section 3 provides an overview of the software and its functionality,
while Section 4 gives an example of how to apply the software to simulated data to illustrate
the correctness of the algorithms. Sections 5 and 6 present two worked examples illustrating
how to use the software for epidemiological and housing market research, while Section 7
concludes with a summary of planned future development for this package.
2. Spatio-temporal models for areal unit data

This section outlines the class of Bayesian hierarchical models that CARBayesST can imple-
ment, and in all cases inference is based on Markov chain Monte Carlo (McMC) simulation.
The first part of this section outlines the general hierarchical model that can be fitted, while
the second describes the range of space-time random effects structures that are available.
2.1. General Bayesian hierarchical model

The study region comprises a set of k = 1, . . . , K non-overlapping areal units S = {S1 , . . . , SK },
and data are recorded for each unit for t = 1, . . . , N consecutive time periods. Thus data are
available for a K N rectangular array with K rows (spatial units) and N columns (time pe-
riods). The response data are denoted by Y = (Y1 , . . . , YN )KN , where Yt = (Y1t , . . . , YKt )
denotes the K 1 column vector of observations for all K spatial units for time period
t. Next, a vector of known offsets are denoted by O = (O1 , . . . , ON )KN , where similarly
Ot = (O1t , . . . , OKt ) denotes the K 1 column vector of offsets for time period t. Finally,
xkt = (xkt1 , . . . , xktp ) is a vector of p known covariates for areal unit k and time period t,
and can include factors or continuous variables and a column of ones for the intercept term.
Note, non-linear covariate-response relationships can be included by adding transformations
of covariates (e.g., squared) or spline basis functions (e.g., using ns()) to the covariate vector.
CARBayesST can fit a generalised linear mixed model to these data, whose general form is:
Ykt |kt f (ykt |kt , 2 ) for k = 1, . . . , K, t = 1, . . . , N, (1)

g(kt ) = x
kt + Okt + kt ,
N( , ).
The vector of covariate regression parameters are denoted by = (1 , . . . , p ), and a multi-

variate Gaussian prior is assumed with mean and diagonal variance matrix that can
be chosen by the user. The kt term is a latent component for areal unit k and time period
t encompassing one or more sets of spatio-temporally autocorrelated random effects, and the
complete set are denoted by = ( 1 , . . . , N ), where t = (1t , . . . , Kt ). CARBayesST
can fit a number of different spatio-temporal structures for kt , all of which are outlined in
Section 2.2 below. The software can implement model (1) for binomial, Gaussian and Poisson
data, and the exact specifications of each are given below:
Binomial - Ykt Binomial(nkt , kt ) and ln(kt /(1 kt )) = x

kt + Okt + kt .
Gaussian - Ykt N(kt , 2 ) and kt = x

kt + Okt + kt .
Poisson - Ykt Poisson(kt ) and ln(kt ) = x

kt + Okt + kt .
In the binomial model (nkt , kt ) respectively denote the number of trials and the probability
of success in each trial in area k and time period t, while in the Gaussian model 2 is the
observation variance. An inverse-gamma(a, b) prior is specified for 2 , and default values of
(a = 1, b = 0.01) are specified by the software unless they are changed by the user.
2.2. Spatio-temporal random effects models

All models implementable in this package induce spatio-temporal autocorrelation into the re-
sponse data Y via the latent random effects component , using CAR type prior distributions
and spatio-temporal extensions thereof. Spatial autocorrelation is controlled by a symmetric
non-negative K K neighbourhood, weight or adjacency matrix W = (wkj ), where wkj rep-
resents the spatial closeness between areal units (Sk , Sj ). Larger valued elements represent
spatial closeness between the two areas in question, where as smaller or zero values corre-
spond to areas that are not spatially close. Typically, W is assumed to be binary, where
wkj = 1 if areal units (Sk , Sj ) share a common border (i.e., are spatially close) and is zero
otherwise. Additionally, wkk = 0. Under this binary specification the values of (kt , jt ) for
spatially adjacent areal units (where wkj = 1) are spatially autocorrelated, where as values
for non-neighbouring areal units (where wkj = 0) are conditionally independent given the
remaining {it } values. This binary specification of W based on sharing a common border is
the most commonly used for areal data, but the only requirement by CARBayesST is for W
to be symmetric, non-negative, and for each row sum to be greater than zero. Similarly, the
model ST.CARanova() defined below uses a binary N N temporal neighbourhood matrix
D = (dtj ), where dtj = 1 if |j t| = 1 and dtj = 0 otherwise. CARBayesST can fit the
following models for :
ST.CARlinear() - fits a model similar to the spatially varying linear time trends model
proposed by Bernardinelli et al. (1995).
ST.CARanova() - fits a model similar to the spatial and temporal main effects and
space-time interaction model proposed by Knorr-Held (2000).
ST.CARsepspatial() - fits a model similar to the overall temporal trend and separate
spatial surfaces model proposed by Napier et al. (2016). Note, this model can only be
applied to binomial or Poisson data.
ST.CARar() - fits the spatially autocorrelated autoregressive time series model of Rush-
worth et al. (2014).
ST.CARadaptive() - fits the localised smoothing model proposed by Rushworth et al.

(2016).
ST.CARlocalised() - fits the localised smoothing and clustering model proposed by

Lee and Lawson (2016). Note, this model can only be applied to binomial or Poisson
data.
This set of models allows users to fit different spatio-temporal structures to their data and
thus examine different underlying hypotheses, and full details of each model are given below.
ST.CARlinear()
This model is a modification of that proposed by Bernardinelli et al. (1995), and estimates
separate linear time trends for each areal unit. Thus it is appropriate if the goal of the
analysis is to estimate which areas are exhibiting increasing or decreasing (linear) trends in
the response over time. The full model specification is given below.
(t t)
kt = 1 + k + ( + k ) , (2)
N
PK !
int j=1 wkj j 2
int
k |k , W N , ,
int K
P PK
j=1 wkj + 1 int int j=1 wkj + 1 int
slo K
P !
j=1 w kj j 2
k | k , W N , PK slo ,
slo K
P
w
j=1 kj + 1 slo slo j=1 wkj + 1 slo
2 2
int , slo Inverse-Gamma(a, b),
int , slo Uniform(0, 1),
N( , 2 ).
Here t = (1/N ) N
P
t=1 t, and thus the modified linear temporal trend covariate is t = (t t)/N
and runs over a centered unit interval. Thus areal unit k has its own linear time trend, with
a spatially varying intercept 1 + k and a spatially varying slope + k . Note, the 1 term
comes from the covariate component x kt in (1). Each set of random effects = (1 , . . . , K )
and = (1 , . . . , K ) are modelled as spatially autocorrelated by Pthe CAR P prior proposed by
Leroux, Lei, and Breslow (2000), and are mean centered, that is K
j=1 j = K
j=1 j = 0. Here
(int , slo ) are spatial dependence parameters, with values of one corresponding to strong spa-
tial smoothness that is equivalent to the intrinsic CAR prior proposed by Besag et al. (1991),
while values of zero correspond to independence (for example if slo = 0 then k N(0, slo 2 )).
Flat uniform priors on the unit interval are specified for the spatial dependence parameters
(int , slo ), while conjugate inverse-gamma and Gaussian priors are specified for the random
effects variances (int 2 , 2 ) and the overall slope parameter respectively. The corresponding
slo
hyperparameters (a, b, , 2 ) can be chosen by the user, and the default values specified by
the software are (a = 1, b = 0.01, = 0, 2 = 1000) which correspond to weakly informative
prior distributions. Alternatively, the dependence parameters (int , slo ) can be fixed at values
in the unit interval [0, 1] rather than being estimated in the model, by specifying arguments
to the ST.CARlinear() function. For example, using the arguments fix.rho.slo=TRUE,
rho.slo=1 sets slo = 1 when fitting the model. Finally, missing (NA) values are allowed in
the response data Y for this model.
ST.CARanova()
The model is a modification of that proposed by Knorr-Held (2000), and decomposes the
spatio-temporal variation in the data into 3 components, an overall spatial effect common
to all time periods, an overall temporal trend common to all spatial units, and a set of
independent space-time interactions. Thus this model is appropriate if the goal is to estimate
overall region wide time trends and spatial patterns. The model specification is given below.
kt = k + t + kt ,
S K
P !
j=1 wkj j S2
k |k , W N , ,
S K
P PK
j=1 wkj + 1 S S j=1 wkj + 1 S
T N
P !
j=1 dtj j T2
t | t , D N , ,
T N
P PN
j=1 dtj + 1 T T j=1 dtj + 1 T
kt N(0, I2 ),
S2 , T2 , I2 Inverse-Gamma(a, b),
S , T Uniform(0, 1).
Here the spatio-temporal autocorrelation is modelled by a common set of spatial random

effects = (1 , . . . , K ) and a common set of temporal random effects = (1 , . . . , N ), and
both are modelled by the CAR prior proposed by Leroux et al. (2000). Additionally, the model
can incorporate an optional set of independent space-time interactions = (11 , . . . , KN ),
which can be specified by the argument interaction=TRUE (the default) in the function call.
All sets of random effects are mean centered. Fixed uniform (S , T ) or conjugate (S2 , T2 , I2 )
priors are specified for the remaining parameters, and the default specifications for the latter
are (a = 1, b = 0.01). Alternatively, in common with the ST.CARlinear() function the
dependence parameters (S , T ) can be fixed at values in the unit interval [0, 1] rather than
being estimated in the model, for full details see the help file for this function. Finally, missing
(NA) values are allowed in the response data Y for this model.
ST.CARsepspatial()
The model is a generalisation of that proposed by Napier et al. (2016), and represents the
data by two components, an overall region-wide temporal trend, and separate spatial surfaces
for each time period which share a common spatial dependence parameter but have different
spatial variances. This model is appropriate if the goal is to estimate both a common overall
temporal trend and the extent to which the spatial variation in the response has changed in
magnitude over time. The model specification is given below.
kt = kt + t ,
S K
P !
j=1 wkj jt t2
kt |kt , W N , ,
S K
P PK
j=1 wkj + 1 S S j=1 wkj + 1 S
T N
P !
j=1 dtj j T2
t | t , D N , ,
T N
P PN
j=1 dtj + 1 T T j=1 dtj + 1 T
2
12 , . . . , N , T2 , Inverse-Gamma(a, b),
S , T Uniform(0, 1),
where k,t = (1,t , . . . , k1,t , k+1,t , . . . , K,t ). This model fits an overall temporal trend
to the data = (1 , . . . , N ) that is common to all areal units, which is augmented with
a separate (uncorrelated) spatial surface t = (1t , . . . , Kt ) at each time period t. The
overall temporal trend and each spatial surface are modelled by the CAR prior proposed by
Leroux et al. (2000), and the latter have a common spatial dependence parameter S but
a temporally-varying variance parameter t2 . Thus the collection (12 , . . . , N 2 ) allows one to
examine the extent to which the magnitude of the spatial variation in the data has changed
over time. Note that here we fix S to be constant in time as it is not orthogonal to t2 ,
thus if it varied then any changes in t2 would not directly correspond to changes in spatial
variance over time. As with all other models the random effects are zero mean centered,
2 ) respectively with
while flat and conjugate priors are specified for (S , T ) and (T2 , 12 , . . . , N
(a = 1, b = 0.01) being the default values. Alternatively, in common with the ST.CARlinear()
function the dependence parameters (S , T ) can be fixed at values in the unit interval [0, 1]
rather than being estimated in the model.
ST.CARar()
The model is that proposed by Rushworth et al. (2014), and models the spatio-temporal
structure as a multivariate first order autoregressive process with a spatially correlated preci-
sion matrix. This model is appropriate if one wishes to estimate the evolution of the spatial
response surface over time without forcing it to be the same for each time period. The model
specification is given below.
kt = kt , (3)
2 1

t |t1 N T t1 , Q(W, S ) t = 2, . . . , N,
1 N 0, 2 Q(W, S )1 ,

2 Inverse-Gamma(a, b),
In this model t = (1t , . . . , Kt ) is the vector of random effects for time period t, which evolve
over time via a multivariate first order autoregressive process with temporal autoregressive
parameter T . The temporal autocorrelation is thus induced via the mean T t1 , while
spatial autocorrelation is induced by the variance 2 Q(W, S )1 . The corresponding precision
matrix Q(W, S ) was proposed by Leroux et al. (2000) and corresponds to the CAR models
used in the other models used above. The algebraic form of this matrix is given by
Q(W, S ) = S [diag(W1) W] + (1 S )I,
where 1 is the K 1 vector of ones while I is the K K identity matrix. In common with all
other models the random effects are zero mean centered, while flat and conjugate priors are
specified for (S , T ) and 2 respectively with (a = 1, b = 0.01) being the default values. In
common with the ST.CARanova() function the dependence parameters (S , T ) can also be
fixed at values in the unit interval [0, 1] rather than being estimated in the model. Finally,
missing (NA) values are allowed in the response data Y for this model.
ST.CARadaptive()
The model is that proposed by Rushworth et al. (2016), and is an extension of ST.CARar()
proposed by Rushworth et al. (2014) to allow for spatially adaptive smoothing. It is appro-
priate if one believes that the residual spatial autocorrelation in the response after accounting
for the covariates is consistent over time but has a localised structure. That is, it is strong in
some parts of the study region but weak in others. The model has the same autoregressive
random effects structure as the previous model ST.CARar(), namely:
kt = kt , (4)
2 1

t |t1 N T t1 , Q(W, S ) t = 2, . . . , N,
1 N 0, 2 Q(W, S )1 ,

However, the random effects from ST.CARar() have a single level of spatial dependence that
is controlled by S . All pairs of adjacent areal units will have strongly autocorrelated random
effects if S is close to one, while no such spatial dependence will exist anywhere if S is close
to zero. However, real data may exhibit spatially varying dependences, as two adjacent areal
units may exhibit similar values suggesting a value of S close to one, while another adjacent
pair may exhibit very different values suggesting a value of S close to zero.
This model allows for localised spatial autocorrelation by allowing spatially neighbouring ran-
dom effects to be correlated (inducing smoothness) or conditionally independent (no smooth-
ing), which is achieved by modelling the non-zero elements of the neighbourhood matrix W
as unknown parameters, rather than assuming they are fixed constants. These adjacency pa-
rameters are collectively denoted by w+ = {wkj |k j}, where k j means areas (k, j) share
a common border. Estimating wkj w+ as equal to zero means (kt , jt ) are conditionally
independent for all time periods t given the remaining random effects, while estimating it close
to one means they are correlated. The adjacency parameters in w+ are each modelled on
the unit interval, by assuming a multivariate Gaussian prior distribution on the transformed
scale v+ = log (w+ /(1 w+ )). This prior is a shrinkage model with a constant mean and
a diagonal variance matrix with variance parameter 2 , and is given by

1 X
f (v+ | 2 , ) exp 2 (vik )2 , (5)
2w +vik v
w2 Inverse-Gamma(a, b).
The prior distribution for v+ assumes that the degree of smoothing between pairs of adjacent
random effects is not spatially dependent, which results from the work of Rushworth et al.
(2016) that shows poor estimation performance when v+ (and hence w+ ) is assumed to
be spatially autocorrelated. Under small values of w2 the elements of v+ are shrunk to ,
and here we follow the work of Rushworth et al. (2016) and fix = 15 because it avoids
numerical issues when transforming between v+ and w+ and implies a prior preference for
values of wkj close to 1. That is as w2 0 the prior becomes the global smoothing model
ST.CARar(), where as when w2 increases both small and large values in w+ are supported
by the prior. As with the other models the default values for the inverse-gamma prior for w2
are (a = 1, b = 0.01). Alternatively, it is possible to fix S using the rhofix argument, e.g.,
rhofix=1 fixes S = 1, so that globally the spatial correlation is strong and is altered locally
by the estimates of w+ . For further details see Rushworth et al. (2016).
ST.CARlocalised()
The model was proposed by Lee and Lawson (2016), and augments the smooth spatio-
temporal variation in ST.CARar() with a piecewise constant intercept component. This
model is appropriate when the aim of the analysis is to identify clusters of areas that ex-
hibit elevated (and reduced) values of the response compared with their geographical and
temporal neighbours. Thus this model is similar to ST.CARadaptive(), in that both relax
the restrictive assumption that if two areas are close together then their estimated random
effects must be similar. This model captures any step-changes in the response via the mean
function, whereas ST.CARadaptive() captures them via the correlation structure (via W).
Model ST.CARlocalised() is given by
kt = Zkt + kt , (6)
2

t |t1 N T t1 , Q(W) t = 2, . . . , N,
1 N 0, 2 Q(W) ,

T Uniform(0, 1).
The random effects = (1 , . . . , T ) are modelled by a simplification of the ST.CARar()

model with S = 1, which corresponds to the intrinsic CAR model proposed by Besag et al.
(1991). Note, for this model the inverse Q(W)1 does not exist as the precision matrix is
singular. This simplification is made so that the random effects capture the globally smooth
spatio-temporal autocorrelation in the data, allowing the other component to capture localised
clustering and step-changes. This second component is a piecewise constant clustering or
intercept component Zkt . Thus spatially and temporally adjacent data points (Ykt , Yjs ) will
be similar (autocorrelated) if they are in the same cluster or intercept, that is if Zkt = Zjs ,
but exhibit a step-change if they are estimated to be in different clusters, that is if Zkt 6= Zjs .
The piecewise constant intercept or clustering component comprises at most G distinct levels,
making this component a piecewise constant intercept term. The G levels are ordered via the
prior specification:
j Uniform(j1 , j+1 ) for j = 1, . . . , G, (7)
where 0 = and G+1 = . Here Zkt {1, . . . , G} and controls the assignment of the
(k, t)th data point to one of the G intercept levels. A penalty based approach is used to model
Zkt , where G is chosen larger than necessary and a penalty prior is used to shrink it to the
middle intercept level. This middle level is G = (G + 1)/2 if G is odd and G = G/2 if G
is even, and this penalty ensures that Zkt is only in the extreme low and high risk classes if
supported by the data. Thus G is the maximum number of distinct intercept terms allowed
in the model, and is not the actual number of intercept terms estimated in the model. The
allocation prior is independent across areal units but correlated in time, and is given by:
exp([(Zkt Zk,t1 )2 + (Zkt G )2 ])

f (Zkt |Zk,t1 ) = PG for t = 2, . . . , N,
2 2
r=1 exp([(r Zk,t1 ) + (r G ) ])
exp((Zk1 G )2 )
f (Zk1 ) =PG ,
2
r=1 exp((r G ) )
Uniform(1, m). (8)
Temporal autocorrelation is induced by the (Zkt Zk,t1 )2 component of the penalty, while
the (Zkt G )2 component penalises class indicators Zkt towards the middle risk class G .
The size of this penalty and hence the amount of smoothing that is imparted on Z is controlled
by , which is assigned a uniform prior. The default value is m = 10, and full details of this
model can be found in Lee and Lawson (2016).
2.3. Inference
All models in this package are fitted in a Bayesian setting using Markov chain Monte Carlo sim-
ulation. A combination of Gibbs sampling (when the appropriate full conditional distributions
are proportional to standard statistical distributions) and Metropolis / Metropolis-Hastings
steps are used, and the majority of the latter use simple random walk proposals. The overall
functions that implement the McMC algorithms are written in R, while the computationally
intensive updating steps are written as computationally efficient C++ routines using the R
package Rcpp (Eddelbuettel and Francois 2011). Additionally, the sparsity of the neighbour-
hood matrices W and D are utilised via their triplet forms when updating the random effects
within the algorithms, which increases the computational efficiency of the software.
3. Loading and using the software
3.1. Loading the software

CARBayesST is a package for the R (R Core Team 2016) statistical software and can be
downloaded from the Comprehensive R Archive Network (CRAN, http://cran.r-project.org/ )
for Windows, Linux and Apple platforms. The package requires R ( 3.0.0) and depends
on packages MASS (Venables and Ripley 2002), and Rcpp ( 0.11.5). Additionally, it im-
ports functionality from the CARBayesdata (Lee 2016), coda (Plummer, Best, Cowles, and
Vines 2006), dplyr (Wickham and Francois 2015) matrixcalc (Novomestky 2012), sp ((Bivand,
Pebesma, and Gomez-Rubio 2013)), spam (Furrer and Sain 2010), spdep, stats, truncdist
(Novomestky and Nadarajah 2012) and utils packages. Once installed it can be loaded using
the command library("CARBayesST").
The packages listed above for use in CARBayesST are automatically loaded by the above call,
but a complete spatial analysis beginning with reading in and formatting shapefiles and data,
creating the neighbourhood matrix W, and plotting the results requires a number of other
packages. Thus the worked examples in this paper utilise functionality from the following
packages: CARBayes, maptools (Bivand and Lewin-Koh 2015), sp and spdep.
3.2. Using the software

The software can fit six models: ST.CARlinear(), ST.CARanova(), ST.CARsepspatial(),
ST.CARar(), ST.CARadaptive() ST.CARlocalised(), and full details of the arguments re-
quired for each function are given in the helpfiles. However, the main arguments common to
all the functions that are required for a baseline analysis (for example using default priors)
are as follows.
formula - A formula for the covariate part of the model using the same syntax used
in the lm() function. Offsets can be included here using the offset() function. The
response and each covariate should be vectors of length KT 1, where each vector is
ordered so that the first K data points are the set of all K spatial locations at time 1,
the next K are the set of spatial points for time 2 and so on.
family - The likelihood model which must be one of binomial, Gaussian or

Poisson.
trials - This is only needed if family=binomial, and is a vector the same length
and in the same order as the response containing the total number of trials for each
area and time period.
W - A K K symmetric and non-negative neighbourhood matrix whose row sums must

all be positive. Typically a binary specification is used, where the kjth element wkj
equals one if areas (Sj , Sk ) are spatially close (e.g., share a common border) and is zero
otherwise. This matrix can be specified by hand or created from a shapefile and data
frame using functionality from the CARBayes and spdep packages.
burnin - The number of McMC samples to discard as the burn-in period.
n.sample - The number of McMC samples to generate.
When a model has run CARBayesST the following summary extractor functions:
coef() - returns the estimated (posterior median) regression coefficients.
fitted() - returns the fitted values based on posterior medians.
logLik() - returns the estimated loglikelihood.
model.matrix() - returns the design matrix of covariates.
print() - prints a summary of the fitted model to the screen, including both parameter
summaries and convergence diagnostics for the McMC run.
residuals() - returns either the response (raw), pearson, or deviance residuals from
the model (based on posterior medians).
Additionally, the CARBayes functions summarise.samples() and summarise.lincomb() can

be applied to CARBayesST models to summarise the results. The software updates the user
on its progress to the R console, which allows the user to monitor the functions progress.
However, using the verbose=FALSE option will disable this feature. Once run, each model in
this package returns a list object with the following components.
summary.results - A summary table of selected parameters that is presented when

using the print() function. The table includes the posterior median (Median) and
95% credible interval (2.5%, 97.5%), the number of samples generated (n.sample), the
acceptance rate for the Markov chain (% accept), the effective number of independent
samples using the effectiveSize() function from the coda package (n.effective),
and the convergence diagnostic proposed by Geweke (1992) and implemented in the
coda package (Geweke.diag). This diagnostic takes the form of a Z-score, so that
convergence is suggested by the statistic being within the range (-1.96, 1.96).
samples - A list containing the McMC samples from the model, where each element in
the list is a matrix. The names of these matrix objects correspond to the parameters
defined in Section 2 of this paper, and each column contains the set of samples for a single
parameter. For example, for ST.CARlinear() the (tau2, rho) elements of the list have
columns ordered as (int2 , 2 ) and (2 , 2 ) respectively. Similarly, for ST.CARanova()
slo int slo
the (tau2, rho) elements of the list have columns ordered as (S2 , T2 , I2 ) (the latter only
if interaction=TRUE) and (2S , 2T ) respectively. Finally, each model returns samples
from the posterior distribution of the fitted values for each data point (fitted).
fitted.values - A vector of fitted values based on posterior medians for each area and
time period in the same order as the data Y.
residuals - A vector of raw residuals based on posterior medians for each area and
time period in the same order as the data Y.
modelfit - Model fit criteria including the Deviance Information Criterion (DIC, Spiegel-
halter, Best, Carlin, and Van der Linde (2002)) and its corresponding estimated effec-
tive number of parameters (p.d), the Watanabe-Akaike Information Criterion (WAIC,
Watanabe (2010)) and its corresponding estimated number of effective parameters (p.w),
and the Log Marginal Predictive Likelihood (LMPL, Congdon (2005)). The best fitting
model is one that minimises the DIC and WAIC but maximises the LMPL.
accept The acceptance probabilities for the parameters.
localised.structure - This element is NULL except for the models ST.CARadaptive()

and ST.CARlocalised(). For ST.CARadaptive() this element is a list with 2 K K
matrices, Wmn and W99 summarising the estimated adjacency relationships. Wmn contains
the posterior median for each wkj element estimated in the model for adjacent areal
units, while W99 contains indicator variables for whether P(wjk < 0.5|Y) > 0.99. For
both matrices, elements corresponding to non-adjacent pairs of areas have NA values.
For ST.CARlocalised() this element is a vector of length KT , and gives the posterior
median class (Zkt value) that each data point is assigned to. This vector is in the same
order as the data Y.
formula - The formula (as a text string) for the covariate and offset part of the model.
model- A text string describing the model that has been fitted.
X - The design matrix of covariates inherited from the formula argument.
The remainder of this paper illustrates the CARBayesST software via a small simulation
study to illustrate the correctness of the McMC algorithms, and two worked examples, the
latter of which utilise spatio-temporal data to answer important questions in public health
and the housing market.
4. Simulation example
This section presents a short simulation example to illustrate both how to use the software
to fit a model and the correctness of the model fitting algorithm. The example here is for
the ST.CARanova() model, and we note that the correctness of the algorithms from the
ST.CARsepspatial() (Napier et al. 2016), ST.CARar() (Rushworth et al. 2014),
ST.CARadaptive() (Rushworth et al. 2016), and ST.CARlocalised() (Lee and Lawson 2016)
models have been assessed in the accompanying papers where the models were developed. We
generate data from a binomial logistic model because the two example analyses that follow in
Sections 5 and 6 utilise Poisson log-linear models. However, similar simulated examples can
easily be created for the other models by adapting the code below.
4.1. Data generation

Consider a spatial region comprising K = 400 areal units on a regular 2020 grid and N = 20
consecutive time periods. Such a grid can be constructed from the code
R> x.easting <- 1:20

R> x.northing <- 1:20
R> Grid <- expand.grid(x.easting, x.northing)
R> K <- nrow(Grid)
R> N <- 20
R> N.all <- N * K
A binary 400400 spatial neighbourhood matrix W can be constructed for this spatial region
based on (rook, in chess) adjacency using the code
R> W <- array(0, c(K, K))

R> for(i in 1:K)
+ {
+ for(j in 1:K)
+ {
+ temp <- (Grid[i, 1] - Grid[j, 1]) ^ 2 +
+ (Grid[i, 2] - Grid[j, 2]) ^ 2
+ if(temp == 1) W[i, j] <- 1
+ }
+ }
whilst a binary 20 20 temporal neighbourhood matrix D can be constructed in a similar

manner using
R> D <- array(0, c(N, N))

R> for(i in 1:N)
+ {
+ for(j in 1:N)
+ {
+ if(abs((i-j)) == 1) D[i, j] <- 1
+ }
+ }
From this the precision matrix can be computed for the multivariate Gaussian representation
of the spatial random effects from (Leroux et al. 2000) as follows:
R> Q.W <- 0.99 * (diag(apply(W, 2, sum)) - W) + 0.01 * diag(rep(1,K))
where S = 0.99. This matrix can then be inverted and a sample of random effects generated
(assuming S2 = 0.01) using the code
R> Q.W.inv <- solve(Q.W)

R> library("MASS")
R> phi <- mvrnorm(n = 1, mu = rep(0, K), Sigma = (0.01 * Q.W.inv))
R> phi.long <- rep(phi, N)
Here the last line repeats the spatial random effects N times, as the ST.CARanova() model
assumes that there is the same set of spatial random effects for each time period. The temporal
random effects under the ST.CARanova() model have the same functional form but depend
on D rather than W, and thus a realisation can be generated analogously using the code
R> Q.D <- 0.99 * (diag(apply(D, 2, sum)) - D) + 0.01 * diag(rep(1, N))

R> Q.D.inv <- solve(Q.D)
R> delta <- mvrnorm(n = 1, mu = rep(0, N), Sigma = (0.01 * Q.D.inv))
R> delta.long <- kronecker(delta, rep(1, K))
Again, the final line repeats the temporal random effects for each spatial unit. Next we
generate space-time interactions and a covariate x, both of which are generated independently
from Gaussian distributions.
R> x <- rnorm(n = N.all, mean = 0, sd = 1)

R> gamma <- rnorm(n = N.all, mean = 0, sd = 0.1)
Finally, assuming an intercept term of 1 = 0.1, a regression coefficient of 2 = 0.1 and the
number of trials for the binomial likelihood in each area and time period being nkt = 50, we
generate the response variable via the code below. Here LP denotes the linear predictor which
contains an intercept term, a covariate and three sets of random effects (spatial, temporal,
and interactions).
R> beta1 <- -0.1

R> beta2 <- 0.1
R> n <- rep(200, N.all)
R> LP <- beta1 + beta2 * x + phi.long + delta.long + gamma
R> theta.true <- exp(LP) / (1 + exp(LP))
R> Y <- rbinom(n = N.all, size = n, prob = theta.true)
4.2. Running the model

The ST.CARanova() model can then be applied to these data using the following code.
R> library("CARBayesST")
R> model <- ST.CARanova(formula = Y~x, family = "binomial", trials = n,
+ W = W, burnin = 20000, n.sample = 220000, thin = 50)
Figure 1: Scatterplots of the true proportions against the estimated proportions from the
model (panel (a)) and the raw sample proportions (panel (b)). The red line is that of equality.
In the code above inference is based on 4,000 MCMC samples, which were generated from a
single Markov chain that was run for 220,000 iterations with a 20,000 burn-in period and sub-
sequently thinned by 50 to remove the temporal autocorrelation. The estimated proportions
kt (posterior medians) are plotted in panel (a) of Figure 1 against the true values theta.true
that generated the data, while panel (b) displays the sample proportions Ykt /nkt against the
true values for comparison purposes. In both cases the red line is the line of equality. The
code to generate the plot is below. As you can see the fitted values are close to the true values
from panel (a), and show markedly less variation than the sample proportions as expected.
R> par(mfrow=c(1, 2))

R> plot(theta.true, model$fitted.values / n, pch = 19, xlab = "True values",
+ ylab = "Estimated proportions from the model", main = "(a)")
R> abline(0, 1, col = "red", lwd = 2)
R> plot(theta.true, Y / n, pch = 19, xlab = "True values",
+ ylab = "sample proportions", main = "(b)")
R> abline(0, 1, col = "red", lwd = 2)
The accuracy of individual components of the model can also be assessed graphically, and
the code below plots the true and estimated values for the spatial random effects surface
. Again, the true and estimated values are close, and the resulting scatterplot is shown in
Figure 2. Note, as the estimated random effects have been mean centred the true values are
also mean centred to aid comparison.
R> par(mfrow = c(1, 1))

R> plot(phi - mean(phi), apply(model$samples$phi, 2, median), pch = 19,
+ xlab = "True values", ylab = "Posterior median")
R> abline(0, 1, col = "red")
Finally, the posterior distributions of the regression parameters (1 , 2 ) can be visualised by

the code below, the result of which is shown in Figure 3.
Figure 2: Scatterplot of the true spatial random effects against the estimates from the model.
The red line is that of equality.
R> colnames(model$samples$beta) <- c("beta1", "beta2")

R> plot(model$samples$beta)
The figures show the Markov chains have converged, and that the posterior distribution of 2
is centred close to the true value of 2 . Note, due to the centering of the random effects in the
model fitting but not in the data generation, the posterior distribution of 1 is not centred
on the true value, as the estimates also incorporates non-zero mean levels from the 3 sets of
random effects. The next two sections give two worked examples of applying the models to
real data to address important questions in epidemiology and housing market research.
5. Example 1 - Quantifying the effect of air pollution on human health

This first example is an ecological regression problem, whose aim is to estimate the effect that
air pollution concentrations have on respiratory disease risk.
5.1. Data and exploratory analysis

For the purposes of delivering health care Scotland is split into 14 health boards, and this
study focuses on the Greater Glasgow and Clyde health board, which contains the city of
Glasgow and has a population of around 1.2 million people during the 2007 to 2011 study
period. This health board is split into K = 271 intermediate geographies (IG), which are
a key geography for the distribution of small-area statistics in Scotland and contain popula-
tions of between 2,468 and 9,517 people. The aim of this study is to quantify the impact of
particulate matter air pollution concentrations on respiratory ill health, and we have yearly
Figure 3: Posterior distributions of the regression parameters (0 , 1 ). The left panel contains
traceplots while the right panel are density estimates.
data for N = 5 years (2007 to 2011) for the K = 271 IGs. The disease and covariate data are
freely available from Scottish Statistics (http://statistics.gov.scot/), while the particu-
late matter pollution concentrations are available from the Department for the Environment,
Food and Rural Affairs (DEFRA, https://uk-air.defra.gov.uk/data/pcm-data).
The respiratory disease data are population level counts of the numbers of admissions to hos-
pital in each IG and year with a primary diagnosis of respiratory disease, which corresponds to
the International Classification of Disease tenth revision (ICD-10) codes J00-J99 and R09.1.
However, the observed numbers of admissions in an IG and year depends on the size and de-
mographic structure (e.g., age and sex profile) of the population living there, which is adjusted
for using indirect standardisation. This involves computing the number of admissions that
would be expected in each IG and year if national age and sex specific admissions rates applied.
The observed and expected numbers of respiratory hospital admissions in the kth IG and tth
year are denoted by (Ykt , Ekt ) respectively, and the Poisson model, Ykt Poisson(Ekt Rkt ) is
typically used to model these data. Here Rkt is the risk, relative to Ekt , of disease in IG k and
year t, and a value of 1.2 corresponds to a 20% increased risk of disease. Operationally, the Ekt
is included as an offset term in the model on the natural log scale, that is Okt = ln(Ekt ) in (1).
The pollution data we utilise are yearly average modelled concentrations of particulate mat-
ter less than 10 microns (PM10 ), which come from both anthropogenic (e.g., particles in car
exhaust fumes) and natural (e.g., sea salt) sources. These data are estimates on a 1 kilome-
tre square grid produced by a numerical simulation model, and full details can be found in
RICARDO-AEA (2015). These 1 kilometre square estimates are spatially misaligned with the
irregularly shaped polygonal IGs at which the disease and covariate data are available, and
thus simple averaging is used to produce IG level PM10 estimates. Specifically, the median
value of PM10 over the set of 1 kilometre grid squares having centroids lying within each IG
was computed, and if an IG was too small to contain a grid square centroid, the nearest grid
square was used as the concentration.
Finally, the data set contains 2 potential confounders that will be included in the model, both
of which are proxy measures of socio-economic deprivation (poverty). The main confounder in
spatio-temporal air pollution and health studies is smoking rates, but such smoking data are
unavailable. However, smoking rates are strongly linked to socio-economic deprivation, and
thus existing studies such as Haining, Li, Maheswaran, Blangiardo, Law, Best, and Richard-
son (2010) control for smoking effects using deprivation based proxy measures. Here we have
two measures of socio-economic deprivation, the average property price in each IG and year
(in hundreds of thousands), and the proportion of the working age population who are in
receipt of job seekers allowance (JSA), the latter being a benefit paid to individuals who are
unemployed and seeking employment.
These data are available in the CARBayesdata package in the object pollutionhealthdata,
and the package also contains the spatial polygon information for the Greater Glasgow and
Clyde health board study region in the object GGHB.IG. These data can be loaded using the
following commands.
R> library("CARBayesdata")
R> library("sp")
R> data("GGHB.IG")
R> data("pollutionhealthdata")
The structure of pollutionhealthdata is shown below
R> head(pollutionhealthdata)
IG year observed expected pm10 jsa price

1 S02000260 2007 97 98.24602 14.02699 2.25 1.150
2 S02000261 2007 15 45.26085 13.30402 0.60 1.640
3 S02000262 2007 49 92.36517 13.30402 0.95 1.750
4 S02000263 2007 44 72.55324 14.00985 0.35 2.385
5 S02000264 2007 68 125.41904 14.08074 0.80 1.645
6 S02000265 2007 24 55.04868 14.08884 1.25 1.760
where the first column labelled IG is the set of unique identifiers for each IG. An exploratory
measure of disease risk is the standardised morbidity ratio (SMR), which for the kth IG and
tth year is computed as SMRkt = Ykt /Ekt . However, due to the natural log link function in
the Poisson model, the covariates are related in the model to the natrual log of the SMR.
Therefore the code below adds the SMR and the natural log of the SMR to the data set and
produces a pairs() plot showing the relationship between the variables.
R> pollutionhealthdata$SMR <- pollutionhealthdata$observed

+ / pollutionhealthdata$expected
R> pollutionhealthdata$logSMR <- log(pollutionhealthdata$SMR)
R> pairs(pollutionhealthdata[ , c(9, 5:7)], pch = 19, cex = 0.5,
+ lower.panel = NULL, labels = c("ln(SMR)", "PM10", "JSA",
+ "Price (*100,000)"))
Figure 4: Scatterplot of the disease, pollution and covariate data.
The pairs plot shown in Figure 4 shows respectively positive and negative relationships be-
tween the natural log of SMR and the two deprivation covariates jsa and price, in both
cases suggesting that increasing levels of poverty are related to an increased risk of respira-
tory hospitalisation. There also appears to be a weak positive relationship between log(SMR)
and PM10 , while the only relationship that exists between the covariates is a negative non-
linear one between jsa and price. Next it is of interest to visualise the average spatial
pattern in the SMR over all five years, and the data can be appropriately aggregated using
the summarise() function from the dplyr package using the code below. The aggregation
is done by the second line, while the final line adds the aggregated averages to the GGHB.IG
SpatialPolygonsDataFrame object.
R> library("dplyr")
R> SMR.av <- summarise(group_by(pollutionhealthdata,IG), SMR.mean =
+ mean(SMR))
R> GGHB.IG@data$SMR <- SMR.av$SMR.mean
Finally, the spplot() function from the sp package can be used to draw a map of the average
SMR over time using the code below:
R> l1 = list("SpatialPolygonsRescale", layout.north.arrow(),

+ offset = c(220000,647000), scale = 4000)
R> l2 = list("SpatialPolygonsRescale", layout.scale.bar(), offset = c(225000,
+ 647000), scale = 10000, fill = c("transparent","black"))
R> l3 = list("sp.text", c(225000,649000), "0")
R> l4 = list("sp.text", c(230000,649000), "5000 m")
R> breakpoints <- seq(min(SMR.av$SMR.mean)-0.1, max(SMR.av$SMR.mean)+0.1,
+ length.out = 11)
R> spplot(GGHB.IG, "SMR", sp.layout = list(l1, l2, l3, l4),
+ col = "transparent", xlab = "Easting", ylab = "Northing",
+ scales = list(draw = TRUE), at = breakpoints,
+ col.regions = terrain.colors(n = length(breakpoints)-1))
Figure 5: Map showing the average SMR over all five years from 2007 to 2011.
The first four lines in the code creates a northarrow and scalebar to add to the map, the fifth
line creates the breakpoints for the colour scheme into 10 equally sized intervals, while the
last line draws the map. The resulting map is shown in Figure 5, where the green shaded
areas are low risk (SMR<1) while the orange to silver areas exhibit elevated risks (SMR>1).
The map shows that the main high-risk areas are in the east-end of Glasgow in the east of
the study region and the Port Glasgow area in the far west of the region on the lower bank of
the river Clyde (the white line running north west to south east). The analysis that follows
requires us to compute the neighbourhood matrix W and a listw object variant of the same
spatial information, the latter being used in a hypothesis test for spatial autocorrelation.
Both of these quantities can be computed from the SpatialPolygonsDataFrame object using
functionality from the spdep package, and code to achieve this is shown below.
R> library("spdep")
R> W.nb <- poly2nb(GGHB.IG, row.names = SMR.av$IG)
R> W.list <- nb2listw(W.nb, style = "B")
R> W <- nb2mat(W.nb, style = "B")
Here W is a binary K K neighbourhood matrix computed based on sharing a common border,

and W.list is the listw object variant of this spatial information.
5.2. Assessing the presence of spatial autocorrelation

The spatio-temporal models in CARBayesST allow for spatio-temporal autocorrelation via
random effects, which capture the remaining autocorrelation in the disease data after the
effects of the known covariates have been accounted for. Therefore, we assess the presence
of spatial autocorrelation in our data set by first computing the residuals from a simple
overdispersed Poisson log-linear model that incorporates the covariate effects. This model is
fitted using the code:
R> formula <- observed ~ offset(log(expected)) + jsa + price + pm10

R> model1 <- glm(formula = formula, family = "quasipoisson",
+ data = pollutionhealthdata)
R> resid.glm <- residuals(model1)
R> summary(model1)$coefficients
Estimate Std. Error t value Pr(>|t|)

(Intercept) -0.59752496 0.054333524 -10.99735 5.287385e-27
jsa 0.06041994 0.003231475 18.69732 1.467196e-69
price -0.28293191 0.018292049 -15.46748 8.225472e-50
pm10 0.04174701 0.003282156 12.71938 4.344434e-35
R> summary(model1)$dispersion
[1] 4.399561
The results show significant effects of all three covariates on disease risk, as well as substantial
overdispersion with respect to the Poisson equal mean and variance assumption (over disper-
sion parameter equal to around 4.40). To quantify the presence of spatial autocorrelation in
the residuals from this model we can compute Morans I statistic (Moran 1950) and conduct
a permutation test for each year of data separately. The permutation test has the null hy-
pothesis of no spatial autocorrelation and an alternative hypothesis of spatial autocorrelation
(either positive or negative), and is conducted using the moran.mc() function from the spdep
package. The test can be implemented for the first year of residuals (2007) using the code
below.
R> moran.mc(x = resid.glm[1:271], listw = W.list, nsim = 10000)
Monte-Carlo simulation of Moran I
data: resid.glm[1:271]
weights: W.list
number of simulations + 1: 10001
statistic = 0.10358, observed rank = 9973, p-value = 0.0028

alternative hypothesis: greater
The estimated Morans I statistic is 0.10358 and the p-value is less than 0.05, suggesting strong
evidence of unexplained spatial autocorrelation in the residuals from 2007 after accounting
for the covariate effects. Similar results were obtained for the other years and are not shown
for brevity. We note that residual temporal autocorrelation could be assessed similarly for
each IG, for example by computing the lag-1 autocorrelation coefficient, but with only 5 time
points the resulting estimates would not be reliable. These results show that the assumption
of independence is not valid for these data, and that spatio-temporal autocorrelation should
be allowed for when estimating the covariate effects.
5.3. Spatio-temporal modelling with CARBayesST

We illustrate model fitting in CARBayesST by applying the ST.CARar() model to the data,
details of which are defined in Section 2. This model has previously been used to account for
spatio-temporal autocorrelation in an air pollution and health study, for details see Rushworth
et al. (2014). The model can be fitted with the following one-line function call, and we note
that all data vectors (response, offset and covariates) have to be ordered so that the first K
data points relate to all spatial units at time 1, the next K data points to all spatial units at
time 2 and so on.
R> model2 <- ST.CARar(formula = formula, family = "poisson",
+ data = pollutionhealthdata, W = W, burnin = 20000, n.sample = 220000,
+ thin = 50)
In the above code the covariate and offset component defined by formula is the same as for
the simple Poisson log-linear model fitted earlier, and the neighbourhood matrix is binary
and defined by whether or not two areas share a common border. The ST.CARar() model is
run for 220,000 McMC samples, the first 20,000 of which are removed by the burn-in period.
The samples are then thinned by 50 to largely remove the autocorrelation, resulting in 4,000
samples for inference. A summary of the model results can be visualised using the print()
function developed for CARBayesST, which gives a very similar summary to that produced
in the CARBayes package.
R> print(model2)
#################
#### Model fitted
#################
Likelihood model - Poisson (log link function)
Latent structure model - Autoregressive CAR model
Regression equation - Y ~ offset(log(E)) + jsa + price + pm10
############
#### Results
############
Posterior quantities for selected parameters and DIC
Median 2.5% 97.5% n.sample % accept n.effective Geweke.diag

(Intercept) -0.6556 -0.8211 -0.4843 4000 30.1 717.9 -1.8
jsa 0.0673 0.0574 0.0770 4000 30.1 887.9 1.0
price -0.1939 -0.2326 -0.1530 4000 30.1 1443.7 1.3
pm10 0.0334 0.0209 0.0444 4000 30.1 613.9 1.5
tau2 0.0591 0.0501 0.0694 4000 100.0 2848.5 0.6
rho.S 0.5701 0.4131 0.7380 4000 45.3 2142.6 1.0

rho.T 0.7555 0.6953 0.8130 4000 100.0 4000.0 0.9
DIC = 10385.78 p.d = 767.0432 LMPL = -4532.71
The Summary is presented in two parts, the first of which describes the model that has been
fit. The second summarises the results, and includes the posterior median (Median) and 95%
credible intervals (2.5%, 97.5%) for selected parameters (not the random effects), the con-
vergence diagnostic proposed by Geweke (1992) (Geweke.diag) as a Z-score and the effective
number of independent samples (n.effective). Also displayed are the actual number of
samples kept from the McMC run (n.sample) as well as the acceptance rate for each param-
eter (% accept). Note, parameters that have an acceptance rate of 100% have been Gibbs
sampled due to their full conditional distributions being a standard distribution. Finally, the
DIC and LMPL overall model fit criteria are displayed, which allows models with different
space-time structures to be compared.
The output from the print() function shows that all three covariates exhibit relationships
with disease risk, as none of the 95% credible intervals contain zero. Furthermore, the spatial
(rho.S) and temporal (rho.T) dependence parameters exhibit relatively high values in the
unit interval, suggesting that both spatial and temporal autocorrelation are present in these
data even after adjusting for the covariate effects. The model object model2 is a list, and
details of its elements are described in Section 3 of this paper. The main components are
vectors of fitted values (model2$fitted.values) and residuals (model2$residuals), and a
list object containing the McMC samples for each individual parameter and the fitted values.
These McMC samples are stored in model2$samples, and each element of the list corresponds
to a different group of parameters. Each element is a mcmc object from the coda package.
Applying the summary() function to this object yields:
R> summary(model2$samples)
Length Class Mode

beta 16000 mcmc numeric
phi 5420000 mcmc numeric
rho 8000 mcmc numeric
tau2 4000 mcmc numeric
fitted 5420000 mcmc numeric
Y 1 mcmc logical
Here the Y object is NA as there are no missing Ykt observations in this data set. If there had
been say m missing values, then the Y component of the list would have contained m columns,
with each one containing posterior predictive samples for one of the missing observations. The
key interest in this analysis is the effects of the covariates on disease risk, which for Poisson
models are typically presented as relative risks. The relative risk for an unit increase in a
covariate with regression parameter 1 is given by the transformation exp(1 ), and a relative
risk of 1.02 corresponds to a 2% increased risk if the covariate increased by . The code below
draws the posterior relative risk distributions for a one unit increase in each covariate, which
are all realistic increases given the variation observed in the data in Figure 4.
Figure 6: Posterior distributions for the covariate effects.
R> colnames(model2$samples$beta) <- c("Intercept", "JSA", "Price", "PM10")

R> plot(exp(model2$samples$beta[ , -1]))
These distributions are displayed in Figure 6, where the left side shows traceplots and the
right side shows density estimates. All three parameters appear to have converged and have
largely symmetric and unimodal posterior distributions. Posterior medians and 95% credible
intervals for the relative risks can be computed using the summarise.samples() function
from the CARBayes package using the code below:
library("CARBayes")
R> parameter.summary <- summarise.samples(exp(model2$samples$beta[ , -1]),
+ quantiles = c(0.5, 0.025, 0.975))
R> round(parameter.summary$quantiles, 3)
0.5 0.025 0.975

[1,] 1.069 1.059 1.080
[2,] 0.823 0.789 0.857
[3,] 1.034 1.021 1.046
The output above shows that the posterior median and 95% credible interval for the relative
risk of a 1gm3 increase in PM10 is 1.034 (1.021, 1.046), suggesting that such an increase
corresponds to 3.4% additional hospital admissions. The corresponding relative risk for a one
percent increase in JSA is 1.069 (1.059, 1.080), while for a one thousand pound increase in
property price (the units for the property price data were in hundreds of thousands) the risk
is 0.823 (0.789, 0.857). Thus, we find that air pollution is related, at this ecological level, to
increased respiratory hospitalisation, while decreased socio-economic deprivation, as measured
by both property price and JSA, is also related to decreased risks of hospital admission.
6. Example 2 - Monitoring the changing state of the housing market

This second example focuses on the state of the housing market, specifically property sales,
and aims to quantify its changing trend over time in an era that encompasses the global
financial crisis that began in late 2007.
6.1. Data and exploratory analysis

The study region is the same as for the first example, namely the set of K = 271 inter-
mediate geographies that make up the Greater Glasgow and Clyde health board. The data
also come from the same source (Scottish Statistics, http://statistics.gov.scot/), and
include yearly observations of house sales from 2003 to 2013 inclusive. The response variable
is the number of property sales Ykt in each IG (indexed by k) and year (indexed by t), and
we additionally have the total number of properties nkt in each IG and year that will be
used in the model as the offset term. We use the following Poisson log-linear model for these
data, Ykt Poisson(nkt kt ), where kt is the rate of property sales as a proportion of the
total number of properties. We note that we have not used a binomial model here as a single
property could sell more than once in a year, meaning that each property does not constitute
a Bernoulli trial. Thus kt is not strictly the proportion of properties that sell in a year, but
is on approximately the same scale for interpretation purposes.
These data are available in the CARBayesdata package in the object salesdata, as is the
spatial polygon information for the Greater Glasgow and Clyde health board study region (in
the object GGHB.IG). These data can be loaded using the following commands.
R> library("CARBayesdata")
R> library("sp")
R> data("GGHB.IG")
R> data("salesdata")
The data.frame salesdata contain 4 columns, the intermediate geography code (IG), the
year the data relate to (year), the number of property sales (sales, Ykt ) and the total number
of properties (stock nkt ). We visualise the temporal trend in these data using the code below,
where the first line creates the raw rate of property sales as a proportion of the total number
of properties.
R> salesdata$salesprop <- salesdata$sales / salesdata$stock

R> boxplot(salesdata$salesprop ~ salesdata$year, range = 0, xlab = "Year",
+ ylab = "Property sales as a proportion of total properties",
+ col = "darkseagreen", border = "navy")
Figure 7: Boxplots showing the temporal trend in the raw rate of property sales as a proportion
of the total number of properties between 2003 and 2013.
This produces the boxplot shown in Figure 7, where the global financial crisis began in 2007.
The plot shows a clear step-change in property sales between 2007 and 2008, as sales were
increasing up to and including 2007, before markedly decreasing in subsequent years. Sales
in the last year of 2013 show slight evidence of increasing relative to the previous 4 years,
possibly suggesting the beginning of an upturn in the market. Also there appears to be a
change in the level of spatial variation from year to year, with larger amounts of spatial
variation observed before the global financial crisis. The spatial pattern in the average (over
time) rate of property sales as a proportion of the total number of properties is shown in
Figure 8, which was created using the code below.
R> library("dplyr")
R> salesprop.av <- summarise(group_by(salesdata, IG),
+ salesprop.mean = mean(salesprop))
R> GGHB.IG@data$sales <- salesprop.av$salesprop.mean
R> l1 = list("SpatialPolygonsRescale", layout.north.arrow(),
+ offset = c(220000,647000), scale = 4000)
R> l2 = list("SpatialPolygonsRescale", layout.scale.bar(),
+ offset = c(225000,647000), scale = 10000,
+ fill = c("transparent","black"))
R> l3 = list("sp.text", c(225000,649000), "0")
R> l4 = list("sp.text", c(230000,649000), "5000 m")
R> breakpoints <- c(0, quantile(salesprop.av$salesprop.mean,
+ seq(0.1, 0.9, 0.1)), 0.1)
R> spplot(GGHB.IG, "sales", sp.layout=list(l1, l2, l3, l4),
Figure 8: Map showing the average (between 2003 to 2013) raw rate of property sales as a
proportion of the total number of properties.

+ col.regions = terrain.colors(n = length(breakpoints)-1))
The first 3 lines of code create the IG specific temporal averages using the dplyr package and
add it to the GGHB.IG SpatialPolygonsDataFrame object, both of which are very similar to
code used in the first example. The remaining lines of code produce the map shown in Figure
8, again using code very similar to that in Example 1. The map and its colour key shows that
the property sales data are skewed to the right, as the color key chosen has unequal groups.
Initially an equally-spaced colour scheme was created, but that showed little colour variation,
hence the use of the unequal quantile based colour key. Secondly, the map shows a largely
similar pattern to that seen for respiratory disease risk in Figure 5, with areas that exhibit
relative high sales rates being the same as those that exhibit relatively low disease risk. Figures
7 and 8 highlight the change in temporal dynamics and the spatial structure in property sales
in Glasgow, and we now apply the ST.CARsepspatial() model from CARBayesST to more
formally quantify these features.
6.2. Quantifying the changing temporal trends and spatial patterns in sales
rates
The extent to which the region-wide average level of sales and its spatial variation and spatial
structure changes over time can be assessed by applying the model proposed by Napier et al.
(2016) to the data, which can be implemented by the ST.CARsepspatial() function. Before
fitting this model we need to create the neighbourhood matrix using the following code:
R> library("spdep")
R> W.nb <- poly2nb(GGHB.IG, row.names = salesprop.av$salesprop.mean)
R> W <- nb2mat(W.nb, style = "B")
Then the model can be fitted using the code below, where inference is again based on 4,000
post burn-in and thinned McMC samples.
R> formula <- sales ~ offset(log(stock))
R> model1 <- ST.CARsepspatial(formula = formula, family = "poisson",
+ data = salesdata, W = W, burnin = 20000, n.sample = 220000, thin = 50)
A summary of the model fit can be obtained using the print() function, and the output is
similar to that shown in example 1 and is not shown for brevity. The model fitted represents
the estimated rate of property sales by
kt = exp(1 + kt + t ),
which is the sum of an overall intercept term 1 , a space-time effect kt with a time period
specific variance, and a region-wide temporal trend t . The mean and standard deviation of
{kt } over space for each year is computed by the following code, which produces the posterior
median and a 95% credible interval for each quantity for each year.
R> trend.mean <- array(NA, c(11, 3))

R> trend.sd <- array(NA, c(11, 3))
R> for(i in 1:11)
+ {
+ posterior <- exp(model1$samples$phi[ , ((i-1) * 271 + 1):(i * 271)] +
+ matrix(rep(model1$samples$beta + model1$samples$delta[ , i], 271),
+ ncol = 271, byrow = FALSE))
+ trend.mean[i, ] <- quantile(apply(posterior, 1, mean),
c(0.5, 0.025, 0.975))
+ trend.sd[i, ] <- quantile(apply(posterior, 1, sd), c(0.5, 0.025, 0.975))
+ }
These temporal trends in the average rate of property sales and its level of spatial variation
can be plotted by the following code, and the result is displayed in Figure 9.
R> par(mfrow=c(2, 1))

R> plot(2003:2013, trend.mean[ , 1], col = "red", type = "l", xlab = "Year",
+ main = "(a)", ylab = "Average property sales rate", ylim = c(0, 0.055))
R> lines(2003:2013, trend.mean[ ,2])
R> lines(2003:2013, trend.mean[ ,3])
R> plot(2003:2013, trend.sd[ ,1], col = "red", type = "l", xlab = "Year",
+ main = "(b)", ylab = "Spatial standard deviation", ylim = c(0, 0.03))
R> lines(2003:2013, trend.sd[ ,2])
R> lines(2003:2013, trend.sd[ ,3])
Figure 9: Posterior median (red) and 95% credible interval (black) for the temporal trend in:
(a) region-wide average property sales rates; and (b) spatial standard deviation in property
sales rates.
The figure shows that both the region-wide average (panel (a)) and the level of spatial varia-
tion (as measured by the spatial standard deviation, panel (b)) in property sales rates show
the same underlying trend, with maximum values just before the global financial crisis in 2007,
and then sharp decreases afterwards. This provides some empirical evidence that the global
financial crisis negatively affected the housing market in Greater Glasgow, with average sales
rates dropping from just under 5.3% in 2007 to 3.3% in 2008. The spatial standard deviation
also dropped from 0.024 to 0.014 over the same two-year period, suggesting that the global
financial crisis had the effect of reducing the disparity in sales rates in different regions of
Glasgow. We note that when measuring the spatial standard deviation we have not simply
presented the posterior distribution of t2 , because this relates to the linear predictor scale
and thus the results change after the inverse logit transformation to the {kt } scale due to a
non-constant mean level (due to the t term).
The posterior median sales rate {kt } is computed and then plotted for the 6 odd numbered
years using the code below, where the colour scale is the same as for Figure 8. The first 3
lines create a data.frame object of estimated sales rates, while the fourth line adds this sales
rate to the SpatialPolygonsDataFrame object. Finally, the last line plots the estimated sales
rates for odd numbered years.
R> rate.est <- matrix(model1$fitted.values / salesdata$stock, nrow = nrow(W),

+ byrow = FALSE)
R> rate.est <- as.data.frame(rate.est)
R> colnames(rate.est) <- c("r2003", "r2004", "r2005", "r2006", "r2007",
+ "r2008", "r2009", "r2010", "r2011", "r2012", "r2013")

R> GGHB.IG@data <- data.frame(GGHB.IG@data, rate.est)
R> spplot(GGHB.IG, c("r2011", "r2013", "r2007", "r2009", "r2003", "r2005"),
+ names.attr = c("Rate 2011", "Rate 2013", "Rate 2007", "Rate 2009",
+ "Rate 2003", "Rate 2005"), sp.layout = list(l1, l2, l3, l4),
+ col.regions = terrain.colors(n = length(breakpoints - 1)))
The maps are displayed in Figure 10, and show the clear changing spatial pattern in sales
rates over time. The spatial rates for 2003 to 2007 are largely consistent, with a correlation
coefficient of 0.76 and a mean absolute difference of 0.012 on approximately (as a property
could sell more than once in a year) the proportion scale. However, a clear step-change is
evident between 2007 and 2009, which incorporates the start of the global financial crisis, with
a correlation coefficient of 0.65 and mean absolute difference of 0.032. The figure shows that
the downturn in sales rates continues into 2011 (the mean absolute difference compared to
2009 is 0.0050) but that the property market is beginning an upturn by 2013. So in conclusion,
Figures 9 and 10 show that the global financial crisis in 2007 resulted in a downturn in both
the regionwide rate of sales and the level of spatial variation in sales across Glasgow, but that
areas of high sales, such as the west-end of Glasgow (the thin strip of orange shaded areas
north of the river in 2009), always remained higher than other parts of the study region.
7. Discussion
This paper has presented the CARBayesST software, which is the first software package dedi-
cated to fitting spatio-temporal CAR type models to areal unit data. This paper has outlined
the class of models that can be implemented by the software together with the Bayesian
inferential framework, as well as the range of specific models that can be fitted. The key ad-
vantage of this package, compared to say implementing the models in BUGS, is its ease of use,
which includes fitting models with a one-line function call, specifying the spatial adjacency
information via a single matrix, and the ability to fit multiple models addressing different
questions about the data in a common software environment. The paper has presented a
small simulation exmple to illustrate the correctness of the algorithms, as well as two fully
worked examples based on quantifying the effect of air pollution on human health and the
changing nature of the housing market.
Future development of the software will be in two main directions. First, as the literature
on spatio-temporal modelling advances we aim to increase the number of spatio-temporal
models that can be implemented, giving the user an even wider set of modelling tools. Second,
with the rapid increase in the availability of small-area data, we aim to develop a suite of
multivariate space-time models (MVST). The development of MVST methodology for areal
unit data is in its infancy, and the ability to jointly examine the spatio-temporal patterns
in multiple response variables simultaneously allows one to address questions that cannot be
addressed by single variable models. For example, in a public health context it allows one to
estimate overall and disease specific spatio-temporal patterns in disease risk, allowing one to
see which areas repeatedly signal at high risk for all diseases and which exhibit elevated risks
Figure 10: Maps showing the changing spatial evolution of the posterior median estimated
sales rates {kt } for odd numbered years.
for only one disease. In the housing context of example 2 a MVST approach would allow one
to extend the analysis carried out by different property types, e.g., flats, terraced houses, etc,
which would allow more insight to be gained about the exact nature of the housing market.
8. Acknowledgements
The Development of the CARBayesST software and this paper were supported by the UK
Engineering and Physical Sciences Research Council (EPSRC, grant EP/J017442/1) and the
UK Medical Research Council (MRC, grant MR/L022184/1). The data and shapefiles used
in this paper were provided by the Scottish Government via their Scottish Statistics website
(http://statistics.gov.scot/).
References
Anderson C, Lee D, Dean N (2014). Identifying Clusters in Bayesian Disease Mapping.

Biostatistics, 15, 457469.
Bakar K, Sahu S (2015). spTimer: Spatio-Temporal Bayesian Modeling Using R. Journal of

Statistical Software, 63(1). URL https://www.jstatsoft.org/index.php/jss/article/
view/v063i15.
Bates D, Machler M, Bolker B, Walker S (2015). Fitting Linear Mixed-Effects Models Using
lme4. Journal of Statistical Software, 67(1). URL https://www.jstatsoft.org/index.
php/jss/article/view/v067i01.
Bernardinelli L, Clayton D, Pascutto C, Montomoli C, Ghislandi M, Songini M (1995).

Bayesian Analysis of Space-Time Variation in Disease Risk. Statistics in Medicine, 14,
24332443.
Besag J, Higdon D (1999). Bayesian Analysis of Agricultural Field Experiments. Journal

of the Royal Statistical Society B, 61, 691746.
Besag J, York J, Mollie A (1991). Bayesian Image Restoration with Two Applications in
Spatial Statistics. Annals of the Institute of Statistics and Mathematics, 43, 159.
Bivand R, Lewin-Koh N (2015). maptools: Tools for Reading and Handling Spatial Objects.
R package version 0.8-36, URL https://CRAN.R-project.org/package=maptools.
Bivand R, Pebesma E, Gomez-Rubio V (2013). Applied Spatial Data Analysis with R. 2nd
edition. Springer-Verlag, New York.
Bivand R, Piras G (2015). Comparing Implementations of Estimation Methods for Spatial

Econometrics. Journal of Statistical Software, 63(18). URL http://www.jstatsoft.org/
v63/i18/.
Charras-Garrido M, Abrial D, de Goer J (2012). Classification Method for Disease Risk

Mapping Based on Discrete Hidden Markov Random Fields. Biostatistics, 13, 241255.
Choi J, Lawson AB (2011). Evaluation of Bayesian Spatial-Temporal Latent Models in Small

Area Health Data. Environmetrics, 22, 10081022.
Congdon P (2005). Bayesian Models for Categorical Data. 1st edition. John Wiley & Sons.
Croissant Y, Millo G (2008). Panel Data Econometrics in R: The plm Package. Journal of
Statistical Software, 27(1). URL https://www.jstatsoft.org/index.php/jss/article/
view/v027i02.
Eddelbuettel D, Francois R (2011). Rcpp: Seamless R and C++ Integration. Journal of

Statistical Software, 40, 8.
Finley A, Banerjee S, Gelfand A (2015). spBayes for Large Univariate and Multivariate
Point-Referenced Spatio-Temporal Data Models. Journal of Statistical Software, 63(1).
URL https://www.jstatsoft.org/index.php/jss/article/view/v063i13.
Furrer R, Sain SR (2010). spam: A Sparse Matrix R Package with Emphasis on MCMC
Methods for Gaussian Markov Random Fields. Journal of Statistical Software, 36(10),
125. URL http://www.jstatsoft.org/v36/i10/.
Geweke J (1992). Evaluating the Accuracy of Sampling-Based Approaches to the Calculation

of Posterior Moments. In Bayesian Statistics, pp. 169193. University Press.
Haining R, Li G, Maheswaran R, Blangiardo M, Law J, Best N, Richardson S (2010). In-

ference from Ecological Models: Estimating the Relative Risk of Stroke from Air Pollution
Exposure Using Small Area Data. Spatial and Spatio-Temporal Epidemiology, 1, 123131,
DOI:10.1016/j.sste.2010.03.006.
Knorr-Held L (2000). Bayesian Modelling of Inseparable Space-Time Variation in Disease

Risk. Statistics in Medicine, 19, 25552567.
Lee D (2013). CARBayes: An R Package for Bayesian Spatial Modelling with Conditional
Autoregressive Priors. Journal of Statistical Software, 55, 13.
Lee D (2016). CARBayesdata: Data Used in the Vignettes Accompanying the CARBayes and
CARBayesST Packages. R package version 2.0.
Lee D, Ferguson C, Mitchell R (2009). Air Pollution and Health in Scotland: A Multicity
Study. Biostatistics, 10, 409423.
Lee D, Lawson A (2016). Quantifying the Spatial Inequality and Temporal Trends in Ma-
ternal Smoking Rates in Glasgow. Annals of Applied Statistics, -, to appear.
Lee D, Minton J, Pryce G (2015). Bayesian Inference for the Dissimilarity Index in the
Presence of Spatial Autocorrelation. Spatial Statistics, 11, 8195.
Leroux BG, Lei X, Breslow N (2000). Statistical Models in Epidemiology, the Environment,
and Clinical Trials, chapter Estimation of Disease Rates in Small Areas: A new Mixed
Model for Spatial Dependence, pp. 179191. Springer-Verlag, New York. URL http:
//dx.doi.org/10.1007/978-1-4612-1284-3_4.
Li G, Best N, Hansell A, Ahmed I, Richardson S (2012). BaySTDetect: Detecting Unusual

Temporal Patterns in Small Area Data via Bayesian Model Choice. Biostatistics, 13,
695710.
Lunn D, Spiegelhalter D, Thomas A, Best N (2009). The BUGS Project: Evolution, Critique
and Future Directions . Statistics in Medicine, 28, 30493082.
Millo G, Piras G (2012). splm: Spatial Panel Data Models in R. Journal of Statistical
Software, 47(1), 138. URL http://www.jstatsoft.org/v47/i01/.
Moran P (1950). Notes on Continuous Stochastic Phenomena. Biometrika, 37, 1723.
Napier G, Lee D, Robertson C, Lawson A, Pollock K (2016). A Model to Estimate the

Impact of Changes in MMR Vaccination Uptake on Inequalities in Measles Susceptibility
in Scotland. Statistical Methods in Medical Research, p. to appear.
Novomestky F (2012). matrixcalc: Collection of Functions for Matrix Calculations. R package

version 1.0-3, URL https://CRAN.R-project.org/package=matrixcalc.
Novomestky F, Nadarajah S (2012). truncdist: Truncated Random Variables. R package

version 1.0-1, URL https://CRAN.R-project.org/package=truncdist.
Paul M, Meyer S (2016). hhh4: An Endemic-Epidemic Modelling Framework for Infectious

Disease Counts, http://CRAN.R-project.org/package=surveillance.
Pinheiro J, Bates D, DebRoy S, Sarkar D, R Core Team (2015). nlme: Linear and Nonlinear
Mixed Effects Models. R package version 3.1-121, URL http://CRAN.R-project.org/
package=nlme.
Plummer M, Best N, Cowles K, Vines K (2006). CODA: Convergence Diagnosis and Out-
put Analysis for MCMC. R News, 6(1), 711. URL http://CRAN.R-project.org/doc/
Rnews/.
R Core Team (2016). R: A Language and Environment for Statistical Computing. R Founda-
tion for Statistical Computing, Vienna, Austria. URL http://www.R-project.org/.
RICARDO-AEA (2015). Technical Report on UK Supplementary Assessment un-

der the Air Quality Directive (2008/50/EC), the Air Quality Framework Directive
(96/62/EC) and Fourth Daughter Directive (2004/107/EC) for 2013. https://uk-
air.defra.gov.uk/data/pcm-data.
Rue H, Martino S, Chopin N (2009). Approximate Bayesian Inference for Latent Gaussian
Models Using Integrated Nested Laplace Approximations (with discussion). Journal of the
Royal Statistical Society B, 71, 319392.
Rushworth A, Lee D, Mitchell R (2014). A Spatio-Temporal Model for Estimating the Long-
Term Effects of Air Pollution on Respiratory Hospital Admissions in Greater London.
Spatial and Spatio-temporal Epidemiology, 10, 2938.
Rushworth A, Lee D, Sarran C (2016). An Adaptive Spatio-Temporal Smoothing Model

for Estimating Trends and Step Changes in Disease Risk. Journal of the Royal Statistical
Society C, -, DOI: 10.1111/rssc.12155.
Schabenberger H (2009). spatcounts: Spatial Count Regression. R package version 1.1, URL
http://CRAN.R-project.org/package=spatcounts.
Spiegelhalter D, Best N, Carlin B, Van der Linde A (2002). Bayesian Measures of Model
Complexity and Fit. Journal of the Royal Statistical Society B, 64, 583639.
Venables WN, Ripley BD (2002). Modern Applied Statistics with S. Fourth edition. Springer-
Verlag, New York. URL http://www.stats.ox.ac.uk/pub/MASS4.
Wakefield J (2007). Disease Mapping and Spatial Regression with Count Data. Biostatistics,
8, 158183.
Wall M (2004). A Close Look at the Spatial Structure Implied by the CAR and SAR Models.
Journal of Statistical Planning and Inference, 121, 311324.
Watanabe S (2010). Asymptotic Equivalence of the Bayes Cross-Validation and Widely Ap-
plicable Information Criterion in Singular Learning Theory. Journal of Machine Learning
Research, 11, 35713594.
Wickham H, Francois R (2015). dplyr: A Grammar of Data Manipulation. R package version

0.4.3, URL https://CRAN.R-project.org/package=dplyr.
Affiliation:
Duncan Lee
School of Mathematics and Statistics
15 University Gardens
University of Glasgow
Glasgow
G12 8QQ, Scotland
E-mail: Duncan.Lee@glasgow.ac.uk
URL: http://www.gla.ac.uk/schools/mathematicsstatistics/staff/duncanlee/
Alastair Rushworth
Mathematics and Statistics
26 Richmond Street
University of Strathclyde
Glasgow
G1 1XH, Scotland
E-mail: alastair.rushworth@strath.ac.uk
URL: https://www.strath.ac.uk/staff/rushworthalastairmr/
Gary Napier
School of Mathematics and Statistics
15 University Gardens
University of Glasgow
Glasgow
G12 8QQ, Scotland
E-mail: Gary.Napier@glasgow.ac.uk
URL: http://www.gla.ac.uk/schools/mathematicsstatistics/staff/garynapier/

CARBayes ST

Hochgeladen von

Dokumentinformationen

Originaltitel

Copyright

Verfügbare Formate

Dieses Dokument teilen

Dokument teilen oder einbetten

Freigabeoptionen

Stufen Sie dieses Dokument als nützlich ein?

Sind diese Inhalte unangemessen?

Copyright:

Verfügbare Formate

CARBayes ST

Hochgeladen von

Copyright:

Verfügbare Formate

Spatio-Temporal Areal Unit Modelling in R with

Conditional Autoregressive Priors Using the

Duncan Lee Alastair Rushworth Gary Napier

Keywords: Bayesian inference, conditional autoregressive priors, R package, spatio-temporal

the Surveillance Epidemiology and End Results programme (http://seer.cancer.gov).

2. Spatio-temporal models for areal unit data

2.1. General Bayesian hierarchical model

Ykt |kt f (ykt |kt , 2 ) for k = 1, . . . , K, t = 1, . . . , N, (1)

The vector of covariate regression parameters are denoted by = (1 , . . . , p ), and a multi-

Binomial - Ykt Binomial(nkt , kt ) and ln(kt /(1 kt )) = x

Gaussian - Ykt N(kt , 2 ) and kt = x

Poisson - Ykt Poisson(kt ) and ln(kt ) = x

2.2. Spatio-temporal random effects models

ST.CARadaptive() - fits the localised smoothing model proposed by Rushworth et al.

ST.CARlocalised() - fits the localised smoothing and clustering model proposed by

Here the spatio-temporal autocorrelation is modelled by a common set of spatial random

Q(W, S ) = S [diag(W1) W] + (1 S )I,

The random effects = (1 , . . . , T ) are modelled by a simplification of the ST.CARar()

j Uniform(j1 , j+1 ) for j = 1, . . . , G, (7)

exp([(Zkt Zk,t1 )2 + (Zkt G )2 ])

3. Loading and using the software

3.1. Loading the software

3.2. Using the software

family - The likelihood model which must be one of binomial, Gaussian or

W - A K K symmetric and non-negative neighbourhood matrix whose row sums must

burnin - The number of McMC samples to discard as the burn-in period.

n.sample - The number of McMC samples to generate.

coef() - returns the estimated (posterior median) regression coefficients.

fitted() - returns the fitted values based on posterior medians.

logLik() - returns the estimated loglikelihood.

model.matrix() - returns the design matrix of covariates.

Additionally, the CARBayes functions summarise.samples() and summarise.lincomb() can

summary.results - A summary table of selected parameters that is presented when

accept The acceptance probabilities for the parameters.

localised.structure - This element is NULL except for the models ST.CARadaptive()

X - The design matrix of covariates inherited from the formula argument.

4.1. Data generation

R> x.easting <- 1:20

R> W <- array(0, c(K, K))

whilst a binary 20 20 temporal neighbourhood matrix D can be constructed in a similar

R> D <- array(0, c(N, N))

R> Q.W <- 0.99 * (diag(apply(W, 2, sum)) - W) + 0.01 * diag(rep(1,K))

R> Q.W.inv <- solve(Q.W)

R> Q.D <- 0.99 * (diag(apply(D, 2, sum)) - D) + 0.01 * diag(rep(1, N))

R> x <- rnorm(n = N.all, mean = 0, sd = 1)

R> beta1 <- -0.1

4.2. Running the model

R> par(mfrow=c(1, 2))

R> par(mfrow = c(1, 1))

Finally, the posterior distributions of the regression parameters (1 , 2 ) can be visualised by

R> colnames(model$samples$beta) <- c("beta1", "beta2")

5. Example 1 - Quantifying the effect of air pollution on human health

5.1. Data and exploratory analysis

The structure of pollutionhealthdata is shown below

IG year observed expected pm10 jsa price

R> pollutionhealthdata$SMR <- pollutionhealthdata$observed

Figure 4: Scatterplot of the disease, pollution and covariate data.

R> l1 = list("SpatialPolygonsRescale", layout.north.arrow(),

Here W is a binary K K neighbourhood matrix computed based on sharing a common border,

5.2. Assessing the presence of spatial autocorrelation

R> formula <- observed ~ offset(log(expected)) + jsa + price + pm10

Estimate Std. Error t value Pr(>|t|)

R> moran.mc(x = resid.glm[1:271], listw = W.list, nsim = 10000)